Due Jun 8, 2:59 AM EDT
The linear regression approximation is
w^=(XTX)−1XTy
Based on the difference in prediction and true value, and the train/test data overall, what ways do you think will help improve the Linear regression model?
Hint: least squares solution tries to aggregate all the given data, even those items that do not seem conform to the general trend. The solution requires finding a matrix inverse, think what properties matrix X has to have.
We won't be able to find the approximation w^, if the matrix XTX is singular.
Considering the multi-feature AlgLinearRegression, which 2 features are the most important, i.e. have the most impact on the prediction?
Use the values in reg.coef_ to determine feature importance. Do not forget, that `coef_` also contains the value for our free variable (intercept).
Use the values in reg.coef_ to determine feature importance. Do not forget, that `coef_` also contains the value for our free variable (intercept).
Consider the multi-feature AlgLinearRegression. What mistakes are there more of, when underpredicting (predicting a value less than the true figure), or overpredicting (predicting a value greater than the true figure)?
Hint: plot a histogram for the error distribution.
The y_pred-y_true distribution is resembling normal distribution.
Looking at the SVM classifier coefficients, which feature is the most important in determining an object class?
Use the "Feature importnaces for SVM" plot.